TransGram for heterogeneity-aware transcriptome reconstruction from long read RNA-Seq data

The advancement of long-read RNA sequencing technologies leads to a bright future for transcriptome analysis. Reconstructing alternative isoforms from long-read RNA data is crucial in transcriptomic studies and poses computational challenges. In this research, we study the heterogeneity of current long reads RNA-seq data, and categorize the sequencing data into two types according to the fraction of the alternative splicing events appeared in the extracted reliably supported sub-path of the splicing graph. Leveraging the heterogeneity, we develop TransGram, an algorithm to reconstruct transcriptome from long RNA-seq reads, which employs the traditional assemble strategy to reconstruct the transcript-representing paths within the splicing graph, and incorporates a machine-learning-based model trained on splice graph-derived features to eliminate false positive transcripts. TransGram supports both annotation-free and annotation-guided modes. Comprehensive evaluations across multiple datasets and conditions demonstrate that TransGram consistently outperforms widely used tools StringTie2, IsoQuant, Bambu, and ESPRESSO, in terms of precision and recall. Long-read RNA sequencing datasets exhibit heterogeneity that affects transcript reconstruction. Here, the authors characterize this heterogeneity and develop TransGram, which adapts reconstruction to different data types to improve accuracy.